Morocco

Multi-task post-training of large language models (LLMs) is typically performed by mixing datasets from different tasks and optimizing them jointly. This approach implicitly assumes that all tasks contribute gradients of similar magnitudes; when this assumption fails, optimization becomes biased toward large-gradient tasks. In this paper, however, we show that this assumption fails in RL post-training: certain tasks produce significantly larger gradients, thus biasing updates toward them. Such gradient imbalance would be justified only if larger gradients implied larger learning gains on the tasks (i.e., larger performance improvements)---but we find this is not true. Large-gradient tasks can achieve similar or even much lower learning gains than small-gradient ones. Further studies reveal that these gradient imbalances cannot be explained by typical training statistics such as training rewards or advantages, suggesting that they are intrinsic to the tasks themselves. This cautions against naive dataset mixing and calls for future work on principled gradient-level corrections for LLMs.

EACL 2026

Imbalanced Gradients in RL Post-Training of Multi-Task LLMs

fine-tuning

reinforcement learning

The purpose of the AAAI conference series is to promote research in Artificial Intelligence (AI) and foster scientific exchange between researchers, practitioners, scientists, students, and engineers across the entirety of AI and its affiliated disciplines. AAAI-23 is the Thirty-Seventh AAAI Conference on Artificial Intelligence. The theme of this conference is to create collaborative bridges within and beyond AI. Like previous AAAI conferences, AAAI-23 will feature technical paper presentations, special tracks, invited speakers, workshops, tutorials, poster sessions, senior member presentations, competitions, and exhibit programs, and two new activities: a Bridge Program and a Lab Program. Many of these activities are tailored to the theme of bridges and all are selected according to the highest standards, with additional programs for students and young researchers. 
AAAI is providing you with a conference planner, which you can use to help organize your itinerary of activities. This includes talks to attend in person, talks to attend remotely, breaks with colleagues and your site seeing activities. To access this conference planner, please go to [https://aaai-2023.takemobi.io](https://aaai-2023.takemobi.io).

In order to access this site, you need to register. If you haven't already, please register [here](https://aaai.org/Conferences/AAAI-23/registration/).


AAAI 2023

The purpose of the AAAI conference series is to promote research in Artificial Intelligence (AI) and foster scientific exchange between researchers, practitioners, scientists, students, and engineers across the entirety of AI and its affiliated disciplines.

technical paper

#### *Message from the General Chair, Aline Villavicencio*
I’m delighted and honoured to welcome you to the 19th Conference of the European Chapter of the Association for Computational Linguistics (EACL 2026), taking place in the beautiful city of Rabat, in Morocco, in March 24-29, 2026. EACL is the flagship European conference of the Association and EACL 2026 proudly continues our field’s tradition of excellence in scholarship, innovation, and inclusivity. I am deeply grateful to the many volunteers whose dedication, generosity, and tireless efforts have made this conference possible.
For the first time EACL is being hosted in the African continent. This is an important milestone for our community, and we are grateful to our Moroccan hosts for enabling this historic moment by bringing this edition of EACL to Rabat. We are also delighted that the Second Arabic NLP School is co-located with EACL. We hope attendees enjoy this wonderful opportunity to strengthen ties with the Computational Linguistics communities across the African continent. *[Read full message](https://drive.google.com/file/d/14NlmHvwM6fPJuMmOvVh7K0vtQbEyv3SZ/view?usp=sharing)* 

<html><button style="display: inline-flex; align-items: center; justify-content: center; white-space: nowrap; border-radius: 9999px; font-weight: bold; background: #7c3aed; color: white; font-family: 'Space Grotesk', sans-serif; height: 40px; font-size: 16px; padding: 0 20px; border: none; cursor: pointer" onclick="window.open('https://underline.io/events/522/reception','_blank')">Go to Workshops and Tutorials Program</button></html>

To access this event page, you need to log in with the **email address you registered with**. Access credentials will be sent to your email from Underline - subject line "Welcome to EACL 2026". Please be sure to check your spam email folder if you do not see an email confirmation right away.

Please log in

To access this event page, you are required to register.
Please complete your registration to continue.

We recommend reading [**the registration information**](https://2026.eacl.org/registration/) first.

**Online Registration Form**: https://acl.swoogo.com/eacl2026

Registration Required

Welcome to the 19th Conference of the European Chapter of the Association for Computational Linguistics (EACL). EACL 2026 will be held in Rabat, Morocco, from March 24–29, 2026. 

As language models evolve into autonomous agents that act and communicate on behalf of users, ensuring safety in multi-agent ecosystems becomes a central challenge. Interactions between personal assistants and external service providers expose a core tension between utility and protection: effective collaboration requires information sharing, yet every exchange creates new attack surfaces. We introduce ConVerse, a dynamic benchmark for evaluating privacy and security risks in agentвЂ“agent interactions. ConVerse spans three practical domains (travel, real estate, insurance) with 12 user personas and over 864 contextually grounded attacks (611 privacy, 253 security). Unlike prior single-agent settings, it models autonomous, multi-turn agent-to-agent conversations where malicious requests are embedded within plausible discourse. Privacy is tested through a three-tier taxonomy assessing abstraction quality, while security attacks target tool use and preference manipulation. Evaluating seven state-of-the-art models reveals persistent vulnerabilitiesвЂ”privacy attacks succeed in up to 88% of cases and security breaches in up to 60%вЂ”with stronger models leaking more. By unifying privacy and security within interactive multi-agent contexts, ConVerse reframes safety as an emergent property of communication.

ConVerse: Benchmarking Contextual Safety in Agent-to-Agent Conversations

Adapting Large Language Models (LLMs) to new domains requires high-quality instruction datasets. Creating these datasets through human annotation is expensive. Current data synthesis methods focus on general tasks and fail to capture domain-specific terminology and reasoning. To address this, we introduce DS -Instruct, a zero-shot framework that generates domain-specific instruction datasets without human supervision. Our approach first generates task-informed keywords to ensure comprehensive domain coverage. It then creates diverse instructions by pairing these keywords with different cognitive levels from Bloom's Taxonomy. Finally, it uses self-consistency validation to filter the outputs and ensure data quality. We apply this framework to generate datasets across seven challenging domains, such as mathematics, finance, and logical reasoning. Thorough evaluations demonstrate that models fine-tuned on our data significantly outperform those obtained from existing methods.

DS2-Instruct: Domain-Specific Data Synthesis for Large Language Models Instruction Tuning

Different large language models (LLMs) exhibit diverse strengths and weaknesses, and LLM ensemble serves as a promising approach to integrate their complementary capabilities. Despite substantial progress in improving ensemble quality, limited attention has been paid to the robustness of ensembles against potential erroneous signals, which often arise from heterogeneous tokenization schemes and varying model expertise. Our analysis shows that ensemble failures typically arise from both the token level and the model level: the former reflects severe disagreement in token predictions, while the latter involves low confidence and pronounced disparities among models. In light of this, we propose CoRE, a plug-and-play technique that harnesses model consistency for robust LLM ensemble, which can be seamlessly integrated with diverse ensemble methods. *Token-level consistency* captures fine-grained disagreements by applying a low-pass filter to downweight uncertain tokens with high inconsistency, often due to token misalignment, thereby improving robustness at a granular level. *Model-level consistency* models global agreement by promoting model outputs with high self-confidence and minimal divergence from others, enhancing robustness at a coarser level. Extensive experiments across diverse benchmarks, model combinations, and ensemble strategies demonstrate that CoRE consistently improves ensemble performance and robustness.

Harnessing Consistency for Robust Test-Time LLM Ensemble

Preference-based alignment like Reinforcement Learning from Human Feedback (RLHF) learns from pairwise preferences, yet the labels are often noisy and inconsistent. Existing uncertainty-aware approaches weight preferences, but ignore a more fundamental factor: the reliability of the answers being compared. To address the problem, we propose Conformal Feedback Alignment (CFA), a framework that grounds preference weighting in the statistical guarantees of Conformal Prediction (CP). CFA quantifies answer-level reliability by constructing conformal prediction sets with controllable coverage and aggregates these reliabilities into principled weights for both DPO- and PPO-style training. Experiments across different datasets show that CFA improves alignment robustness and data efficiency, highlighting that modeling answer-side uncertainty complements preference-level weighting and yields more robust, data-efficient alignment.

Conformal Feedback Alignment: Quantifying Answer-Level Reliability for Robust LLM Alignment

Illicit drug use among teenagers and young adults (TYAs) remains a pressing public health concern, with rising prevalence and long-term impacts on health and well-being. To detect illicit drug use among TYAs, researchers analyze large-scale surveys such as the Youth Risk Behavior Survey (YRBS) and the National Survey on Drug Use and Health (NSDUH), which preserve rich demographic, psychological, and environmental factors related to substance use. However, existing modeling methods treat survey variables independently, overlooking latent and interconnected structures among them. To address this limitation, we propose LAMI (LAtent relation Mining with bi-modal Interpretability), a novel joint graph-language modeling framework for detecting illicit drug use and interpreting behavioral risk factors among TYAs. LAMI represents individual responses as relational graphs, learns latent connections through a specialized graph structure learning layer, and integrates a large language model to generate natural language explanations grounded in both graph structures and survey semantics. Experiments on the YRBS and NSDUH datasets show that LAMI outperforms competitive baselines in predictive accuracy. Interpretability analyses further demonstrate that LAMI reveals meaningful behavioral substructures and psychosocial pathways, such as family dynamics, peer influence, and school-related distress, that align with established risk factors for substance use. Our codebase is available here.

Interpretable Graph-Language Modeling for Detecting Youth Illicit Drug Use

Large language models have shown strong reasoning capabilities through chain-structured methods such as Chain-of-Thought. Recent studies optimize thought structures by generating parallel or tree-like structures, switching long and short reasoning modes, or aligning reasoning steps with task performance. However, these approaches mainly rely on previously generated logical directions of the chains, which ignore the unexplored regions of the solution space. Such a phenomenon is denoted as blind spots, which limit the diversity and effectiveness of the reasoning process. To this end, we propose the ``Thought Space Explorer'' (TSE), a framework for navigating and expanding thought structures to overcome blind spots in LLM reasoning. Our TSE first identifies key nodes with high impact, then generates new nodes by integrating information from multiple chains. Finally, it extends new branches through connection strategies. We conduct a series of experiments on math and QA benchmarks. Compared to existing baseline methods, TSE improves the accuracy of both the final answer and intermediate reasoning steps, while maintaining a better effectiveness-efficiency trade-off for practical deployment.

Blind Spot Navigation in Large Language Model Reasoning with Thought Space Explorer

As the scale of Large Language Models (LLMs) continues to grow rapidly, the cost of training and inference has significantly increased, limiting their application in resource-constrained scenarios. To address this challenge, model pruning has been widely used to reduce computational complexity. Among various pruning strategies, block-wise pruning has gained popularity due to its ability to accelerate computation by removing entire blocks of parameters. However, existing methods often rely on hard labels from calibration datasets and neglect the cumulative effects of pruning on subsequent blocks. To address this, we propose two complementary techniques: the Logit Disruption Score (LDS), a novel block importance criterion that measures the impact of pruning by comparing the cosine similarity between the logits of the original and pruned models, focusing on the most informative logit dimensions to better preserve the model's core capabilities; and Activation Statistics Correction (ASC), an affine transformation mechanism that aligns the mean and variance of activations in the pruned model with those of the original model, effectively mitigating the distribution shift caused by block removal and improving the information flow in subsequent blocks. Experiments across multiple datasets show that our approach reduces reliance on calibration data and improves generalization, achieving competitive results with existing methods.

Logits-Based Block Pruning with Affine Transformations for Large Language Models

The inherent risk of generating harmful and unsafe content by Large Language Models (LLMs), has highlighted the need for their safety alignment. Various techniques like supervised fine-tuning, reinforcement learning from human feedback, and red-teaming were developed for ensuring the safety alignment of LLMs. However, the robustness of these aligned LLMs is always challenged by adversarial attacks that exploit unexplored and underlying vulnerabilities of the safety alignment. In this paper, we develop a novel black-box jailbreak attack, called BitBypass, that leverages hyphen-separated bitstream camouflage for jailbreaking aligned LLMs. This represents a new direction in jailbreaking by exploiting fundamental information representation of data as continuous bits, rather than leveraging prompt engineering or adversarial manipulations. Our evaluation of five state-of-the-art LLMs, namely GPT-4o, Gemini 1.5, Claude 3.5, Llama 3.1, and Mixtral, in adversarial perspective, revealed the capabilities of BitBypass in bypassing their safety alignment and tricking them into generating harmful and unsafe content. Further, we observed that BitBypass outperforms several state-of-the-art jailbreak attacks in terms of stealthiness and attack success. Overall, these results highlights the effectiveness and efficiency of BitBypass in jailbreaking these state-of-the-art LLMs.

BitBypass: A New Direction in Jailbreaking Aligned Large Language Models with Bitstream Camouflage

Steering vectors are a lightweight method for controlling text properties by adding a learned bias to language model activations at inference time. While predominantly studied for multiple-choice and toy tasks, their effectiveness in free-form generation remains largely unexplored. Moving ``Beyond Multiple Choice,'' we thoroughly evaluate the effectiveness of steering vectors in adaptively controlling topical focus, sentiment, toxicity, and readability in abstractive summaries from the SAMSum, NEWTS and arXiv datasets. We find that steering effectively controls the targeted summary properties, but high steering strengths consistently degrade both intrinsic and extrinsic text quality. Prompting alone offers weaker control, while preserving text quality. Combining steering and prompting yields the strongest control over text properties and offers the most favorable efficacy-quality trade-off at moderate steering strengths. Our results show steering vectors are an efficient method for controlling summarization, but introduce a critical trade-off between control strength and text quality. We demonstrate this balance can be optimized through hybrid techniques, which offer the most favorable efficacy-quality trade-off.

Beyond Multiple Choice: Evaluating Steering Vectors for Summarization

Identifying LLM-generated code through watermarking poses a challenge in preserving functional correctness. Previous methods rely on the assumption that watermarking high-entropy tokens effectively maintains output quality. Our analysis reveals a fundamental limitation of this assumption: syntax-critical tokens such as keywords often exhibit the highest entropy, making existing approaches vulnerable to logic corruption. We present STONE, a syntax-aware watermarking method that embeds watermarks only in non-syntactic tokens and preserves code integrity. For its rigorous assessment, we also introduce STEM, a comprehensive framework that balances three critical dimensions: correctness, detectability, and imperceptibility. Across Python, C++, and Java, STONE preserves correctness, sustains strong detectability, and achieves balanced performance with minimal overhead. Our implementation is available at https://anonymous.4open.science/r/STONE-watermarking-AB4B/.

Downloads

Next from EACL 2026

ConVerse: Benchmarking Contextual Safety in Agent-to-Agent Conversations

Similar lecture

Continual Learning with Scaled Gradient Projection

Stay up to date with the latest Underline news!

PRODUCT

COMPANY

RESOURCES